E-commerce
June 30, 2026
"Is this jacket really in excellent condition?" "Is the stain in photo 3 acceptable for a Very Good rating?" "Do you have the same one in size 38?" On a second-hand product page, the customer is not comparing two identical SKUs: they are evaluating a unique item. A standard new catalogue bot will reply "48h delivery" when the question is about actual wear and tear and stock alternatives.
Faume observes a direct correlation between the visual richness of the pre-owned product page and the add-to-cart rate: the more photos and measured details there are, the closer the conversion rate gets to the standard new-clothing experience (Faume, second-hand conversion 2026). Congruence Market Insights estimates that 58% of leading resale platforms already use computer vision for grading, but pre-purchase dialogue remains under-automated (Congruence, resale tech 2026).
This guide #287 covers the recommerce AI chatbot: explaining individual conditions and defects, and proposing listing-to-listing alternatives. It complements second-hand support (#286) (human SNAD macros) and refurbished bot (#270) (standardised A/B/C grades) with a focus on decision support for unique items.
Summary
Why does a generic bot fail at recommerce?
The recommerce chatbot must speak in terms of listing_id, individual photos, and listed flaws, not homogeneous stock variants.
Three generic bot failures
Generic condition: "good condition" without citing the grade or the defects of the viewed listing
Promise of new: minimizing wear ("imperceptible") outside of the listing data
Blind alternatives: suggesting size S when the customer wants the same model in size 38
Recommerce buyer expectations
MDPI study on 524 US consumers: information accuracy and authenticity are the two main drivers of trust before a resale purchase (MDPI, trust circular fashion 2025). The bot must respond like an expert seller who has seen the item, not like a brand-new product help center knowledge base.
Automatable volume
Brand take-back store or certified marketplace: 45 to 65% of pre-purchase questions = grade, visible defect, measurements, authenticity badge, same model alternative. Naratix estimates that listings with explicit defect photos reduce "is this authentic?" messages (Naratix, secondhand catalog 2026).
How does it differ from refurbished (#270) and support (#286)?
Five neighboring pieces of content, five bot roles.
Second-hand human support (#286)
Second-hand (#286): SNAD protocol, HS macros, dispute photos. The #287: tier 1 bot tier, defects, alternatives before escalation.
Refurbished Bot (#270)
Refurbished (#270): A/B/C grades, battery %, workshop tests. The #287: Excellent to Fair grading, patina, unique item without reset.
Out-of-stock alternatives (#OOS)
OOS alternatives bot: new SKU substitute in the same range. The #287: another recommerce listing (higher grade, similar size, same brand).
Product comparison
Comparison bot: catalog specs. The #287: compare two listings with different defects and prices.
Test grid (#283)
Test grid bot (#283): generic QA. The #287 provides a pack of 30 recommerce questions listing-specific.
Which intent matrix to automate vs. handoff?
The recommerce bot intents matrix separates reliable answers on unit sheets and dispute files.
Auto (listing context + confidence ≥ 95 %)
recom_grade_explain: in-house definition of Excellent/Very good/Good/Fairrecom_defect_detail: readdefects_list[]+ photo redirect Nrecom_measurements: measurements in cm from sheetrecom_auth_badge:auth_verifiedstatus + daterecom_return_policy: final sale or 14 days if applicablerecom_photos_gallery: listing angle linksrecom_alternative_same_model: other listings with the same parent SKUrecom_alternative_grade_up: same size, higher graderecom_alternative_size: same model, size ±1 if in stockrecom_vs_new: % savings vs new + expected wear
Immediate Handoff
recom_snad_open: dispute for item not as describedrecom_auth_dispute: suspected counterfeit post-purchaserecom_p2p_mediation: third-party seller marketplacerecom_exception_final_sale: return request outside policy
Golden rule
The bot always cites the consulted listing_id. Never describe the condition of another listing without an explicit change of context.
How to configure voice and guardrails on a single room?
The recommerce bot voice is transparent about wear and tear, never apologetic about second-hand.
System prompt principles
Formal address (vouvoiement), reassuring, factual tone
Cite the condition grade + each listed defect before any positive adjective
Refer to the photo number for each defect ("see photo 4")
Never say "as new" unless grade is Excellent + defects_list is empty
Forbidden guardrails
Inventing a defect absent from
defects_listMinimizing a stain/hole ("barely visible") not confirmed by the listing
Guaranteeing authenticity beyond the displayed badge
Promising SNAD refund without workflow #286
Offering an alternative whose
listing_idis sold or reserved
AI Transparency
Recommerce PDP greeting: "AI Assistant. I only answer inquiries regarding this item #[listing_id]. For any dispute after reception, I will put you in touch with an advisor."
What is the seven-step pre-purchase flow on the listing PDP?
The recommerce bot pre-purchase flow on the PDP listing converts doubt into an informed decision.
Contextual welcome: detect recommerce collection + inject session listing_id
Open-ended question: "What would you like to check on this item?"
Grade + defects: read condition_grade + defects_list, photo links
Measurements / fit: measurements_cm vs brand size guide
Authenticity: auth_verified badge or process if pending
Return: final_sale or withdrawal based on policy
Alternative if hesitating: suggest max 2 other listings with supporting arguments
Example jacket Very good
Customer: "Is the brand mark on the collar readable?" Bot: "Listing #SH-8842, grade Very good. Listed defects: slight patina on the inside collar (photo 5), no exterior stains. Embroidered logo readable on photo 2. Measurements: shoulders 41 cm, length 62 cm."
Budget hesitation
Customer: "A bit expensive for second-hand." Bot: "Same model #SH-9011 in Good condition, −28%, visible elbow wear in photo 3. Or #SH-8755 Very good condition, −12%, with no listed defects."
Which Shopify metafields populate the RAG corpus?
The recommerce RAG bot corpus combines help center #286 and unit listing metafields.
Required Shopify Metafields
recom.listing_id: unique identifierrecom.condition_grade: Excellent, Very good, Good, Fairrecom.defects_list: JSON [{type, location, photo_index}]recom.measurements_cm: shoulders, chest size, length…recom.auth_status: verified | pending | nonerecom.auth_date: ISO date if verifiedrecom.final_sale: booleanrecom.parent_sku: link to new model for price comparisonrecom.photos_annotated: URLs + defect indexes
Help center chunks
Pages: grading grid #286, authenticity process, SNAD returns summary. Tag vertical:recommerce for filtered retrieval.
Sync webhook
Update defects_list or grade → re-index listing chunk within 15 min. Discrepancy between page / bot = post-delivery SNAD spike.
Alternative options index
Secondary table: parent_sku + size + condition_grade + price + availability. Query bot before any alternative recommendation.
How do I explain each defect with a photo reference?
Explaining listing defects by photo is the core differentiator of the recommerce bot.
Defect response template
Structure: [Grade] + [Defect type + location] + [Photo index] + [Impact when worn]. Example: "Good: light wear on left lapel, visible in photo 4 in natural light. Does not affect fit or closure."
ACM Research on marketplaces
A 2025 ACM study of 929 users confirms that second-hand buyers judge image quality also on defect visibility, not just overall aesthetics (ACM, resale image quality 2025). The bot systematically points to the defect-on-photo macro, never to a vague promise.
Frequently asked questions about defects
"Stain in photo 3, washable?": quote defects_list; if not specified → "not tested for cleaning, sold as is"
"Hole repaired?": read repair type in defects_list
"Odor?": smell field if present; otherwise "visual inspection only"
"Normal wear for the grade?": compare defect vs grade definition section 3 #286
Prohibited
"You won't notice it once worn" without listing basis. Authorized phrasing: "Classified as Good: expected wear for this grade."
How to suggest listing-to-listing alternatives?
The recommerce alternatives bot engine replaces an item, not a new out-of-stock SKU.
Recommendation tree (max 2 suggestions)
Same parent_sku + same size: higher grade if customer budget is mentioned
Same parent_sku + size ±1: if customer measurements or account history allow
Same brand + category: if exact model is sold out
New parent_sku: last resort with price delta and "new" argument
Suggestion format
"Alternative 1: #SH-9011, Good, €89 (−28% vs current listing), wear on elbows photo 3. Alternative 2: #SH-8755, Very good, €112, no listed defect." Direct PDP listing link.
Single stock guardrails
Verify availability=available and reserved_until null before recommending. Item in another session's cart: exclude or display "reserved 15 min".
Side-by-side comparison
Intent recom_compare_listings: chat table showing grade | price | defects | auth | final_sale for 2 listing_id max.
New listing alert
If no alternatives are available: capture email + parent_sku + size + minimum desired grade. Notify when matching listing is published.
Where to place human handoff on SNAD and authenticity?
The recommerce handoff bot protects the brand on disputes and authenticity.
Auto-escalation signals
Words: counterfeit, fake, scam, lawyer, fraud authorities
Post-purchase + SNAD keywords
3 turns without resolution on the same defect
Request for additional photo outside the gallery (ops only)
Gorgias handoff payload
listing_id, grade, defects_list JSON, transcript of the last 5 messages, PDP photo URLs, proposed alternatives, auth_status.
Limited post-purchase bot
In-house WISMO inventory OK. SNAD opening: collect pack photos #286 section 7 then escalate, never promise a refund.
P2P marketplace
Neutral bot, no seller favoritism. Mediation intent → trust team with escrow status.
Which KPIs measure the impact of the recommerce bot?
Measure the recommerce bot ROI separately from new and refurbished items.
Conversation tags
recom_grade, recom_defect, recom_auth, recom_alt_proposed, recom_alt_clicked, recom_snad_escalated, recom_final_sale_question.
Monthly KPIs
Pre-purchase self-resolution: % sessions without handoff (target 35-50%)
Assisted conversion: chat listing_id orders / recommerce bot PDP sessions
Alternative accept rate: alternative listing purchase / proposals
Post-bot SNAD rate: disputes / bot-assisted orders (must decrease vs non-assisted)
Time-to-decision: median minutes session → add-to-cart
Weekly review
Top 10 unanswered RAG questions → enrich defects_list or help. Cross-reference with the friction report (#281).
How does Qstomy automate assistance in choosing recommerce?
Qstomy injects the recommerce listing context into bot and handoff without state hallucination.
Capabilities
Intent recom_*: routing section 3
Lookup listing_id: metafields + indexed photo gallery
Alternatives engine: parent_sku, grade, size, live stock
Guardrail defects: blocks response outside defects_list
Enriched Handoff: payload section 9 to Gorgias/Zendesk
Compare 2 listings in-chat
Encrypted DTC Scenario
Fashion trade-in brand (in-house inventory), 420 active listings, 2,800 PDP recommerce sessions/month, auto-resolution 19%, SNAD 7.1% of orders. Deployment of Qstomy recom flows + metafields + alternatives. After 8 weeks: auto-resolution 44%, assisted conversion +31%, alternative accept 18%, SNAD 7.1 → 3.4%, human pre-purchase tickets −52%.
Explore Shopify, AI support, demo.
Which playbooks should be used to deploy the recommerce bot?
Playbook 1: metafields listing (2 d)
Deploy section 6 fields on all recommerce listings. Backfill 50 pilot listings. QA bot on 10 detail pages.
Playbook 2: intents + prompt (1 d)
Configure section 3 matrix, section 4 system prompt, guardrails. Shadow mode 1 week.
Playbook 3: 7-step PDP flow (4 h)
Activate section 5 flow on recommerce collection. Trigger button "Ask a question about this item".
Playbook 4: alternatives index (1 d)
Build parent_sku + availability table. Test section 8 tree on 20 size/grade/budget scenarios.
Playbook 5: 30-question test pack (3 h)
Grade, photo defect, auth, final sale, alternative, compare, simulated SNAD. Grid #283 adapted for recommerce.
Useful linking
This week: open 5 bot conversations on real recommerce listings. Does the bot mention the listing_id and the photo of the defect? If not, correct the corpus before expanding traffic.
Enzo
June 30, 2026
